A Byte of Coding Issue 401

Howdy,

Hope you all had a nice weekend. I spent most of mine fiddling around with modelling electricity prices for individual providers in the US (for findenergy.com). The issue is we get monthly sales / revenues / customers figures for some big providers in every state, but the rest are all reported anonymously in one block (more than half of the state total). All providers report on an annual basis, but that’s usually 1-2 years behind, compared to monthly data which is only 3 months behind. Obviously the latter is more interesting for most people.

My theory was that if I could find the proportion of total annual sales / revenues / customers for each provider that published monthly data anonymously, I could use that ratio to predict what proportion of the monthly anonymous data could be attributed to them. The result works super well! Except for 2020. Maybe because of the pandemic? For some reason the model amount of electricity sold is much lower than the actual electricity sold, whereas revenues and customers are pretty much on point. Maybe a combination of more electricity use with subsidies that aren’t counted as revenue? What do you think?

Anyway, here’s the issue.

Made possible through generous sponsorship by:

How Good Are the Latest Open LLMs? And Is DPO Better Than PPO?

Published: 12 May 2024

Tags: llm, ai, machine learning

Sebastian Raschka evaluates recent advancements in large language models (LLMs) such as Mixtral, Llama 3, and Phi-3, and discusses reinforcement learning methods for model fine-tuning, comparing Proximal Policy Optimization (PPO) and Direct Preference Optimization (DPO).

Some highlights:

• Compares newer models to previous versions, highlighting improvements in architecture and training strategies.

• Discusses the efficiency and scalability of models, particularly in terms of data use and processing capabilities.

• Reviews practical implications of PPO and DPO in aligning LLMs with human preferences.

When Is a Counter-Strike Player Good?

Published: 10 May 2024

Tags: math, data science, games

Chris reflects on the complexities of measuring skill in Counter-Strike, highlighting the multifaceted nature of in-game performance which extends beyond simple metrics like kills and deaths, focusing instead on strategic gameplay and resource management.

Some highlights:

• Discusses the Leetify rating, a sophisticated metric that evaluates individual performance based on the impact of in-game events on the chances of winning.

• Provides insights into the tactical considerations of Counter-Strike, such as map control and economy management.

• Explores the difficulty of assessing individual contributions to team success in a game where strategic decision-making is crucial.

Single Sign-On: Acronyms Demystified

Published: 1 June 2023

Tags: sponsored, auth

WorkOS explains all the auth-related protocols and provides additional resources (for further reading) for OAuth, OIDC, and SAML.

Some highlights:

• “OAuth is not an authentication protocol”

• OIDC is built on top of OAuth

• SAML is used for federated authentication

Playing with Zig and thinking about its package manager

Published: 12 April 2024

Tags: zig

Paul Hammant explores using the Zig programming language to build a virtual machine configuration tool and discusses the potential of federated Git repositories as a decentralized package management system.

Some highlights:

• Demonstrates a practical application of Zig by creating a command-line tool for virtual machine configuration.

• Proposes leveraging Git for decentralized package management, akin to systems used by Carthage and Ansible Galaxy.

• Discusses the benefits of a Git-based system for robust version control and distribution of software binaries.